Method and apparatus for generating recommendation information

Abstract

The present disclosure provides a method and a corresponding device for generating recommendation information. The method includes: determining a conversion relationship graph based on conversion data corresponding to a recommended item recommended by a recommender, where the conversion relationship graph includes a plurality of recommender nodes and a plurality of recommended item nodes, and an edge connecting a recommender node and a recommended item node indicates that the recommender has recommended the recommended item; determining a set of recommendation relationship sequences including a plurality of recommendation relationship sequences based on the conversion relationship graph, where each recommendation relationship sequence includes a plurality of alternating recommender identifiers and recommended item identifiers; generating at least one of a recommender vector and a recommended item vector based on the set of recommendation relationship sequences; and generating recommendation information to be delivered based on the similarity between at least one of the recommender vector and the recommended item vector. The present disclosure can expand the recommendation scope of products, improve the product selection efficiency, and promote the increase in transaction volume and transaction efficiency.

Description

Technical Field

[0001] The present disclosure relates to the field of multimedia, and more particularly, to a method, apparatus, device, computer-readable storage medium, and computer program product for generating recommendation information. Background Art

[0002] Currently, there has emerged an advertising promotion model of the CPS (Cost per Sales) alliance type. The CPS alliance calculates advertising fees based on the actual number of products sold, which can most directly reflect the advertising promotion effect. In the CPS scenario, Key Opinion Consumers (KOCs) earn commissions by promoting products and according to the actual sales situation of the products. Key Opinion Consumers, also known as influencers, generally refer to influential consumers on the Internet, who are often followed by a large number of ordinary consumers. After an influencer promotes a product, ordinary consumers can purchase the product through the link shared by the influencer.

[0003] However, influencers also need to select products to be recommended / promoted from a vast number of products, and this process is also called product selection. Currently, influencers often select products based on their own history and transaction effects. Since such a product selection method is only based on the influencer's own historical behavior, it is easy to form an information cocoon, resulting in the dilemma of limited recommended product scope and slow improvement of transaction capabilities. Therefore, there is a need to provide a method for generating recommendation information to further expand the scope of recommended products and improve transaction efficiency. Summary of the Invention

[0004] To solve the above problems, the present disclosure provides a method, apparatus, device, computer-readable storage medium, and computer program product for generating recommendation information.

[0005] According to one aspect of the embodiments of the present disclosure, a method for generating recommendation information is provided, including: determining a conversion relationship graph based on conversion data corresponding to a recommended item recommended by a recommender, the conversion relationship graph including a plurality of recommender nodes and a plurality of recommended item nodes, an edge connecting a recommender node and a recommended item node indicating that the recommender has recommended the recommended item, the recommender node being identified by a recommender identifier, and the recommended item node being identified by a recommended item identifier; determining a set of recommendation relationship sequences including a plurality of recommendation relationship sequences based on the conversion relationship graph, each recommendation relationship sequence including a plurality of alternating recommender identifiers and recommended item identifiers, wherein, in each recommendation relationship sequence, the recommender corresponding to the recommender identifier has recommended the recommended item corresponding to the recommended item identifier adjacent to the recommender identifier; generating at least one of a recommender vector and a recommended item vector based on the set of recommendation relationship sequences; and generating recommendation information to be placed based on the similarity between at least one of the recommender vector and the recommended item vector.

[0006] For example, the conversion relationship graph includes a plurality of basic conversion relationship units, and each basic conversion relationship unit includes at least: an identifier of a recommender, an identifier of a recommended item, and a weight corresponding to an edge connecting the recommender node and the recommended item node, where the weight indicates a combination of one or more of the following: the value of the recommended item, the number of purchases or the total purchase value generated based on the recommendation of the recommended item by the recommender.

[0007] For example, determining the conversion relationship graph based on the conversion data corresponding to the recommender's recommendation of the recommended item further includes: aggregating the recommendation data of the recommender's recommendation of the recommended item and the transaction data of the purchaser's purchase of the recommended item to generate the conversion data corresponding to the recommender's recommendation of the recommended item, where the purchaser purchases the recommended item at least partially based on the recommendation information of the recommended item by the recommender; and determining the conversion relationship graph based on the conversion data corresponding to the recommender's recommendation of the recommended item.

[0008] For example, aggregating the recommendation records of the recommender's recommendation of the recommended item and the transaction records of the purchaser's purchase of the recommended item further includes: obtaining a plurality of recommendation records based on the recommendation data of the recommender's recommendation of the recommended item, where each recommendation record includes a recommender identifier, a recommended item identifier, and a recommendation event identifier; obtaining a plurality of transaction records based on the transaction data of the purchaser's purchase of the recommended item, where each transaction record includes a purchaser identifier, a recommended item identifier, and a recommendation event identifier; aggregating the plurality of recommendation records and the plurality of transaction records with the recommendation event identifier as the aggregation dimension to generate a plurality of conversion records, where each conversion record includes a recommender identifier, a recommended item identifier, and a purchaser identifier; and integrating the plurality of conversion records into the conversion data corresponding to the recommender's recommendation of the recommended item.

[0009] For example, determining the conversion relationship graph based on the conversion data corresponding to the recommender's recommendation of the recommended item further includes: traversing each recommendation record in the recommendation data of the recommender's recommendation of the recommended item to determine the edge connecting the recommender node and the recommended item node in the conversion relationship graph; traversing each transaction record in the transaction data of the purchaser's purchase of the recommended item, and determining the weight corresponding to the edge connecting the recommender node and the recommended item node based on the conversion data corresponding to the recommender's recommendation of the recommended item; and matching the recommender node, the recommended item node, and the weight corresponding to the edge connecting the recommender node and the recommended item node to generate a basic conversion relationship unit in the conversion relationship graph.

[0010] For example, determining the conversion relationship graph based on the conversion data corresponding to the recommended items by the recommender further includes: obtaining a first basic conversion relationship unit not incorporated into the conversion relationship graph, and determining a data shard corresponding to the first basic conversion relationship unit, where the data shard is used to store a sub-graph of the conversion relationship graph; and updating the corresponding data shard by using the first basic conversion relationship unit.

[0011] For example, the updating the corresponding data shard by using the first basic conversion relationship unit further includes: obtaining the recommender identifier, the recommended item identifier, and the weight corresponding to the edge connecting the recommender node and the recommended item node in the first basic conversion relationship unit; determining whether the sub-graph of the conversion relationship graph stored in the data shard includes a second basic conversion relationship unit having the same recommender identifier and recommended item identifier as those of the first basic conversion relationship unit based on the recommender identifier and the recommended item identifier in the first basic conversion relationship unit; updating the weight in the second basic conversion relationship unit based on the weight in the first basic conversion relationship unit when the sub-graph of the conversion relationship graph stored in the data shard includes the second basic conversion relationship unit; and storing the first basic conversion relationship unit when the sub-graph of the conversion relationship graph stored in the data shard does not include the second basic conversion relationship unit.

[0012] For example, determining the recommended relationship sequence set including a plurality of recommended relationship sequences based on the conversion relationship graph further includes: determining a starting element corresponding to each recommended relationship sequence in the recommended relationship sequence set based on the conversion relationship graph, where the starting element is a recommender identifier or a recommended item identifier; and randomly walking starting from the starting element corresponding to each recommended relationship sequence to generate each recommended relationship sequence in the recommended relationship sequence set, and the length of each recommended relationship sequence is less than a preset maximum length.

[0013] For example, the randomly walking to generate each recommended relationship sequence in the recommended relationship sequence set further includes: for each recommended relationship sequence, adding the starting element corresponding to the recommended relationship sequence to the recommended relationship sequence, and using the starting element as the current element, determining the next element of the current element based on the depth walking parameter, the breadth walking parameter, and the weights corresponding to the respective edges connecting the current element, where the next element is different from the elements already added to the recommended relationship sequence; and adding the next element to the recommended relationship sequence, and using the next element as the current element until the length of the recommended relationship sequence is equal to the preset maximum length or the next element is empty.

[0014] For example, generating at least one of a recommended direction vector and a recommended item vector based on the set of recommendation relationship sequences further includes: for each recommending party node or each recommended item node in the conversion relationship graph, using a word vector conversion model to generate the recommended direction vector corresponding to the recommending party or the recommended item vector corresponding to the recommended item, where the word vector conversion model is trained by the set of recommendation relationship sequences, and the distance between two recommended direction vectors with higher similarity is smaller, and the distance between two recommended item vectors with higher similarity is smaller.

[0015] For example, the training of the word vector conversion model includes: using each recommending party identifier and each recommended item identifier in the conversion relationship graph as central words respectively, calculating the co-occurrence vector matrix corresponding to the central words based on the set of recommendation relationship sequences, where the co-occurrence vector indicates the conditional probability distribution of the corresponding relationship between the central word and the neighborhood of the central word in the conversion relationship graph, determining the predicted co-occurrence vector matrix corresponding to the central word based on the central word, using the dictionary central word vector matrix and the dictionary context word vector matrix in the word vector conversion model, and adjusting the dictionary central word vector matrix and the dictionary context word vector matrix so that the predicted co-occurrence vector matrix corresponding to the central word approximates the co-occurrence vector matrix corresponding to the central word.

[0016] For example, generating the recommended information to be delivered based on the similarity between at least one of the recommended direction vector and the recommended item vector further includes: determining a recommending party identifier or a recommended item identifier based on the scenario information corresponding to the recommended information to be delivered; determining the recommended direction vector or the recommended item vector corresponding to the recommending party identifier or the recommended item identifier based on the recommending party identifier or the recommended item identifier, and using the recommended direction vector or the recommended item vector as a query key; using the query key to retrieve multiple vectors similar to the recommended direction vector or the recommended item vector from a retrieval database, where the multiple vectors are a combination of multiple recommended direction vectors, multiple recommended item vectors, or recommended direction vectors and recommended item vectors, and generating the recommended information to be delivered based on the multiple vectors; where the retrieval database includes at least one of multiple recommending party entries and multiple recommended item entries, where each recommending party entry includes: a recommending party identifier and the recommended direction vector corresponding to the recommending party identifier, and each recommended item entry in the retrieval database includes: a recommended item identifier and the recommended item vector corresponding to the recommended item identifier.

[0017] For example, in the case where the scenario information indicates recommending alternative recommended items to the purchaser based on a key recommender, the recommender identifier is the recommender identifier corresponding to the key recommender; in the case where the scenario information indicates recommending alternative recommenders similar to the key recommender to the key recommender, the recommender identifier is the recommender identifier corresponding to the key recommender; in the case where the scenario information indicates recommending alternative recommended items to the key recommender, the recommender identifier is the recommender identifier corresponding to the key recommender; in the case where the scenario information indicates recommending alternative recommended items to the purchaser based on the recommended items previously purchased by the purchaser, the recommended item identifier is the recommended item identifier corresponding to the recommended items previously purchased by the purchaser; or in the case where the scenario information indicates recommending alternative recommended items to the purchaser or the key recommender based on the currently viewed recommended item, the recommended item identifier is the recommended item identifier corresponding to the currently viewed recommended item.

[0018] For example, generating the recommended information to be delivered based on the multiple recommender vectors or recommended item vectors further includes one or more of the following: in the case where the scenario information indicates recommending alternative recommended items to the purchaser based on a key recommender, generating the first recommended information to be delivered based on the recommender vector corresponding to the key recommender identifier, where the first recommended information includes information on multiple alternative recommended items associated with the key recommender; in the case where the scenario information indicates recommending alternative recommenders similar to the key recommender to the key recommender, generating the second recommended information to be delivered based on the recommender vector corresponding to the key recommender identifier, where the second recommended information includes information on multiple alternative recommenders similar to the key recommender; in the case where the scenario information indicates recommending alternative recommended items to the key recommender, generating the third recommended information to be delivered based on the recommender vector corresponding to the key recommender, where the third recommended information includes information on multiple alternative recommended items associated with the key recommender; in the case where the scenario information indicates recommending alternative recommended items to the purchaser based on the recommended items previously purchased by the purchaser, generating the fourth recommended information to be delivered based on the recommended item vector corresponding to the recommended items previously purchased by the purchaser, where the fourth recommended information includes information on multiple alternative recommended items associated with the recommended items previously purchased by the purchaser; or in the case where the scenario information indicates recommending alternative recommended items to the purchaser or the key recommender based on the currently viewed recommended item, generating the fifth recommended information to be delivered based on the recommended item vector corresponding to the currently viewed recommended item, where the fifth recommended information includes information on multiple alternative recommended items associated with the currently viewed recommended item.

[0019] According to another aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium having computer-readable instructions stored thereon, and when the computer-readable instructions are executed by a processor, the processor is caused to execute the method according to any one of the above aspects of the present disclosure.

[0020] According to another aspect of the embodiments of the present disclosure, there is provided a computer program product, which includes computer-readable instructions, and when the computer-readable instructions are executed by a processor, the processor is caused to execute the method according to any one of the above aspects of the present disclosure.

[0021] By using the method, device, equipment, computer-readable storage medium, and computer program product for generating recommendation information according to the above aspects of the present disclosure, it is possible to construct a conversion relationship graph between recommenders and recommended items, and based on this conversion relationship graph, assist the recommender in product selection or the purchaser in purchasing goods, thereby expanding the product recommendation scope, improving the product selection efficiency, and promoting the increase in transaction volume and transaction efficiency. In addition, the present disclosure also improves the calculation efficiency based on large-scale graph storage technology and efficient graph computing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] By describing the embodiments of the present disclosure in more detail in conjunction with the accompanying drawings, the above and other objects, features, and advantages of the embodiments of the present disclosure will become more obvious. The accompanying drawings are used to provide a further understanding of the embodiments of the present disclosure, and constitute a part of the specification, and are used to explain the present disclosure together with the embodiments of the present disclosure, and do not constitute a limitation to the present disclosure. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0023] Figure 1 A schematic diagram showing an application scenario according to an embodiment of the present disclosure is shown.

[0024] Figure 2 A flowchart showing the method for generating recommendation information according to an embodiment of the present disclosure is shown.

[0025] Figure 3 A schematic diagram showing the method for generating recommendation information according to an embodiment of the present disclosure is shown.

[0026] Figure 4 An example diagram showing partial transaction data according to an embodiment of the present disclosure is shown.

[0027] Figure 5 An example diagram showing partial recommendation data according to an embodiment of the present disclosure is shown.

[0028] Figure 6 An example diagram showing partial conversion data according to an embodiment of the present disclosure is shown.

[0029] Figure 7It is an example diagram showing the relationship among a recommender - a recommended item - a purchaser according to an embodiment of the present disclosure.

[0030] Figure 8 It is an example diagram showing the conversion relationship diagram between a recommender and a recommended item according to an embodiment of the present disclosure.

[0031] Figure 9 It is a partial example diagram showing the overall business diagram according to an embodiment of the present disclosure.

[0032] Figure 10 It shows a flowchart of storing and updating the conversion relationship diagram according to an embodiment of the present disclosure.

[0033] Figure 11 It is an example diagram showing the determination of a set of recommended relationship sequences according to an embodiment of the present disclosure.

[0034] Figure 12 It is an example diagram showing the process of training a word vector conversion model according to an embodiment of the present disclosure.

[0035] Figure 13 It is an example diagram showing the process of retrieving similar recommended direction vectors and recommended item vectors according to an embodiment of the present disclosure.

[0036] Figure 14 It shows a schematic diagram of the architecture of an exemplary computing device according to an embodiment of the present disclosure. Detailed implementation manners

[0037] Next, the technical solutions in the embodiments of the present disclosure will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present disclosure.

[0038] Embodiments of the present disclosure may be based on artificial intelligence (AI). Artificial intelligence is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. For example, for the embodiments of the present disclosure, it can recommend a certain recommended item to a user in a way similar to how human judgment selects products of interest to a talent / user from a vast amount and recommends them to the talent / user. Artificial intelligence enables the embodiments of the present disclosure to have the functions of understanding user behavior and reasoning and processing user preferences by studying the design principles and implementation methods of various intelligent machines. Artificial intelligence technology covers a wide range of fields, including both hardware-level technologies and software-level technologies. Among them, artificial intelligence software technologies mainly include computer vision technology, natural language processing, and machine learning / deep learning, etc.

[0039] In addition, embodiments of the present disclosure also relate to cloud computing technology. Cloud computing is a computing model that distributes computing tasks (for example, calculating user preferences for each of multiple scenarios) on a resource pool composed of a large number of computing devices, enabling various application systems to obtain computing power, storage space, and information services as needed. The network that provides resources is called the "cloud". The resources in the "cloud" seem to be infinitely expandable to users, and can be obtained at any time, used on demand, expanded at any time, and paid according to usage.

[0040] The present disclosure provides a method for generating recommendation information, as well as corresponding devices and equipment. Embodiments of the present disclosure can construct a conversion relationship graph between a recommender and a recommended item, and based on this conversion relationship graph, assist the recommender in product selection or the purchaser in purchasing goods, thereby expanding the product recommendation scope, improving the product selection efficiency, and promoting the increase in transaction volume and transaction efficiency. In addition, the present disclosure also improves the computing efficiency based on large-scale graph storage technology and efficient graph computing technology.

[0041] First, refer to Figure 1 Describe the application scenarios of the method for generating recommendation information and the corresponding devices, etc. according to the embodiments of the present disclosure. Figure 1 FIG. 13 shows a schematic diagram of application scenario 100 according to an embodiment of the present disclosure, in which a server 110 and multiple terminals 120 are schematically shown.

[0042] According to the method for generating recommendation information and the corresponding device of the embodiment of the present disclosure, the server 110 can be mounted on the server 110 to generate recommendation information to be delivered. The server 110 here can be an independent server for generating recommendation information, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, positioning services, and big data and artificial intelligence platforms. The embodiment of the present disclosure does not make specific restrictions on this. Each of the multiple terminals 120 can be a fixed terminal such as a desktop computer, a mobile terminal such as a smart phone, a tablet computer, a portable computer, a handheld device, a personal digital assistant, a smart wearable device, or any combination thereof, and the embodiment of the present disclosure does not make specific restrictions on this.

[0043] Optionally, Figure 1 One or more of the multiple terminals 120 can serve as recommenders, and other terminals 120 can serve as buyers. The server 110 can serve as a promotion platform, which is used to deliver recommendation information to recommenders or buyers for display on the terminals 120 corresponding to the recommenders or buyers. For example, the recommendation information can be information in the form of pictures, texts, videos, or any combination thereof. At present, the server 110 often generates recommendation information to assist the recommender in selecting products based on the recommender's own product selection history and transaction results, or generates recommendation information to assist the audience in purchasing products based on the audience's purchase history. Because such a method of generating recommendation information is often prone to forming an information cocoon, it results in the dilemma of limited recommended product range and slow improvement in transaction capabilities. To this end, the present disclosure proposes a method for generating recommendation information, which can further expand the scope of recommended products and improve transaction efficiency.

[0044] Refer to the following Figures 2 to 3 A method for generating recommendation information according to an embodiment of the present disclosure is described. Figure 2 A flowchart of a method 200 for generating recommendation information according to an embodiment of the present disclosure is shown. Figure 3 A schematic diagram of a method 200 for generating recommendation information according to an embodiment of the present disclosure is shown.

[0045] like Figure 2 As shown, the method 200 for generating recommendation information according to an embodiment of the present disclosure exemplarily includes steps S210 to S240, but the present disclosure is not limited thereto. Steps S210 to S230 are also referred to as offline steps of the method 200, which can be performed when the server 110 is offline or when the server 110 is online. Step S240 is also referred to as an online step of the method 200, which is performed when the server 110 is online.

[0046] Optionally, the offline step is used to integrate the original data stored in the background (e.g., transaction data and recommendation data) through big data processing frameworks such as Spark, Hadoop, and MPI to obtain the metadata of the transformed data graph. Then, the metadata is imported into the graph storage engine and batch processed to generate the training samples required to solve the recommendation direction vector and the recommended item vector. Then, the vector representations of each recommender / recommended item in the transformed relationship graph are obtained using a machine learning framework, and these vector representations are stored in the retrieval library.

[0047] Optionally, the online step is used to: obtain the information of the current recommender / purchaser / recommended item, and retrieve how many approximate recommended items, multiple approximate purchasers, or multiple approximate recommenders in the vector retrieval library based on one or more of the recommendation direction vector, the purchase direction vector, and the recommended item vector. Then, the retrieval results are filtered and sorted and presented to the recommender (influencer) or the audience. The influencer selects products and promotes them to the audience. At the same time, the online step records the recommendation data of the recommender and the transaction information of the purchaser for the offline step to further update the transformed data, the transformed relationship graph, the recommendation relationship sequence set, and the recommendation direction vector or the recommended item vector.

[0048] Next, continue to refer to Figure 2 and Figure 3 to further describe steps S210 to S240.

[0049] First, in step S210, based on the transformed data corresponding to the recommender's recommendation of the recommended item, a transformed relationship graph is determined. The transformed relationship graph includes multiple recommender nodes and multiple recommended item nodes. The edge connecting the recommender node and the recommended item node indicates that the recommender has recommended the recommended item. The recommender node is identified by a recommender identifier, and the recommended item node is identified by a recommended item identifier.

[0050] For example, a recommender generally refers to a key opinion consumer (influencer) or an advertiser. A promotion platform refers to a party that uses its own platform or technology to help the recommender select products, analyze, predict, etc., and generates recommendation information. The audience refers to the user group that browses the recommendation information, and they often follow one or more recommenders (e.g., influencers) through social media and consumption platforms. An audience who has purchased the recommended item recommended by the recommender is also called a purchaser.

[0051] Refer to Figure 3, for example, in the example scenario of the recommender's product selection, the recommendation information can be any information that helps the recommender select products. In this scenario, the server 110 generates recommendation information to assist the recommender in product selection. Then, after the recommender selects products, a recommended item link will be published, and this process will generate recommendation data. Subsequently, the audience following the recommender may click on the recommended item link. For example, they can jump to the product page for browsing by clicking on the product link in the advertisement. After the audience clicks and jumps to the product page, they may purchase the product on the product page, or download and install the application on the product page, etc. The audience who purchases the product or downloads the application is then referred to as the purchaser, and the actions of purchasing the product and downloading the application will generate purchase data. The process of transforming from an audience browsing the product link to a purchaser will correspondingly generate conversion data. The conversion data can be derived from the above-mentioned recommendation data and transaction data.

[0052] Also, for example, in the scenario of the purchaser buying a product, the recommendation information can be any information that helps the purchaser select a product. The server 110 generates recommendation information to assist the audience in purchasing the product. The server 110 directly delivers these recommendation information to the audience for easy purchase, and this process also generates recommendation data. Correspondingly, the audience's purchase of the product will generate transaction data. The process of the audience transforming from a reader of the recommendation information to a purchaser will also correspondingly generate conversion data. The present disclosure does not further limit the delivery process of the recommendation information.

[0053] The database stores the recommendation data of the recommender's recommended items and the transaction data of the purchaser's purchase of the recommended items. Since the recommendation data and the transaction data of the purchaser's purchase of the recommended items are often massive, the embodiments of the present disclosure can further adopt big data technology. For example, through big data processing frameworks such as Spark, Hadoop, and MPI, the recommendation data and the transaction data can be integrated to obtain the conversion data. Big data refers to a collection of data that cannot be captured, managed, and processed by conventional software tools within a certain time range, and is a massive, high-growth, and diverse information asset that requires new processing models to have stronger decision-making power, insight discovery power, and process optimization capabilities. With the advent of the cloud era, big data has attracted more and more attention. Based on the fact that big data requires special technologies to effectively implement the media information processing method provided in this embodiment, the technologies applicable to big data include massively parallel processing databases, data mining, distributed file systems, distributed databases, and cloud computing, etc. The conversion data corresponding to the recommender's recommended items can be derived at least based on the recommendation data of the recommender's recommended items and the transaction data of the purchaser's purchase of the recommended items. Then, this process will be described with reference to Figures 4 to 6 and will not be elaborated here.

[0054] Continue to refer to Figure 3, the above conversion data reflects the trading volume and / or transaction amount generated based on the recommendation of the recommended item by the recommender. That is, the conversion data reflects the correlation between the recommender, the recommended item, and the purchaser at the numerical level. Furthermore, a conversion relationship graph can be constructed using the conversion data. For example, the conversion relationship graph presents the correlation between the recommender, the purchaser, and the recommended item in the form of the data structure of a graph. The conversion relationship graph is, for example, an undirected graph. It includes a finite (possibly variable) set as the node set, and a set of unordered pairs (corresponding to the undirected graph) as the edge set. In addition, the conversion relationship graph can also be a directed graph, and the present disclosure is not limited thereto.

[0055] A node can be a part of the conversion relationship graph structure or an external entity represented by an integer subscript or reference. The data structure of the conversion relationship graph may also include a numerical value (edge value) associated with each edge, such as a label or a numerical value (i.e., weight). For example, the node set in the conversion relationship graph can include any one or more of the nodes representing the recommended item or the recommended items. The edge connecting the recommender node and the recommended item node indicates that the recommender recommends the recommended item, and the weight of the edge indicates the trading volume or transaction amount generated based on the recommender's recommendation of the recommended item, etc. In addition, the node set in the conversion relationship graph can also include a purchaser node, and the edge connecting the purchaser and the recommended item indicates that the purchaser has purchased the recommended item. The edge connecting the purchaser and the recommender indicates that the purchaser has followed the recommender.

[0056] Optionally, the conversion relationship graph can be split into multiple subgraphs and stored distributively. The subgraph that cannot be further split will be used as the basic unit constituting the conversion relationship graph, and will be referred to as the basic conversion relationship unit hereinafter. As an example, the conversion relationship graph includes multiple basic conversion relationship units, and each basic conversion relationship unit includes at least: the identifier of the recommender, the identifier of the recommended item, and the weight corresponding to the edge connecting the recommender node and the recommended item node, and the weight indicates one or more combinations of the following: the value of the recommended item, the purchase quantity or total purchase value generated based on the recommender's recommendation of the recommended item. Thereafter, reference will be made to Figures 7 to 9 For further description of the example of the conversion relationship graph, the present disclosure will not elaborate herein. Those skilled in the art should understand that the composition methods of the conversion relationship graphs describing different types of nodes will vary slightly, and the present disclosure does not impose any restrictions on the composition method and the expression method of the conversion relationship graph.

[0057] For example, the determined conversion relationship graph can be stored in a memory in the form of an adjacency list, an adjacency matrix, or an incidence matrix, etc. However, since the conversion data is massive and changes in real time, the conversion relationship graph can also be stored in a distributed memory. Thereafter, reference will be made to Figure 10An example of storing and updating the transformation relationship graph is further described, which will not be elaborated in this disclosure. This disclosure does not limit the specific storage form and storage structure of the transformation relationship graph.

[0058] Next, in step S220, based on the transformation relationship graph, a set of recommendation relationship sequences including multiple recommendation relationship sequences is determined, and each recommendation relationship sequence includes multiple alternating recommender identifiers and recommended item identifiers. Among them, in each of the recommendation relationship sequences, the recommender corresponding to the recommender identifier recommends the recommended item corresponding to the recommended item identifier adjacent to the recommender identifier. For example, the transformation relationship graph includes a large number of nodes and edges, and the transformation relationship graph may also change in real time according to the behaviors of recommenders and purchasers. It is almost impossible to traverse all the nodes and edges in the transformation relationship graph. Therefore, it is necessary to sample the data in the transformation relationship graph to save the amount of calculation. To further improve the operation efficiency, a sampling method combined with batch processing technology can also be adopted. The sampling methods include Struc2Vec, DeepWalk, Node2Vec, etc. This disclosure will refer to Figure 11 The process of sampling the transformation relationship graph in the Node2Vec manner to generate a set of recommendation relationship sequences is further described, which will not be elaborated here.

[0059] Next, in step S230, at least one of a recommender vector and a recommended item vector is generated based on the set of recommendation relationship sequences. For example, a machine learning solution can be adopted to generate at least one of a recommender vector and a recommended item vector based on the set of recommendation relationship sequences. With the development of machine learning, various neural network models can be used to complete the above-mentioned machine learning tasks. For example, a deep neural network (DNN) model, a factorization machine (FM) model, etc. can be adopted. These neural network models can be implemented as acyclic graphs, where neurons are arranged in different layers. Generally, a neural network model includes an input layer and an output layer, and the input layer and the output layer are separated by at least one hidden layer. The hidden layer transforms the input received by the input layer into a representation useful for generating an output in the output layer. The network nodes are fully connected to the nodes in the adjacent layer via edges, and there are no edges between the nodes within each layer. The data received at the nodes of the input layer of the neural network is propagated to the nodes of the output layer via any one of a hidden layer, an activation layer, a pooling layer, a convolutional layer, etc. The input and output of the neural network model can adopt various forms, and this disclosure does not limit this.

[0060] For example, the recommender vector is used to characterize the association relationship and approximate relationship between the recommender and other nodes in the data space of the transformation relationship graph. The recommended item vector is used to characterize the association relationship and approximate relationship between the recommended item and other nodes in the data space of the transformation relationship graph. Optionally, methods such as Skip-Gram can be used to obtain the above-mentioned recommender vector and recommended item vector. This disclosure will refer toFigure 12 Examples of generating the recommendation direction vector and the recommended item vector are further described, which will not be elaborated here.

[0061] Next, in step S240, based on the similarity between at least one of the recommendation direction vector and the recommended item vector, the recommendation information to be delivered is generated.

[0062] For example, the similarity of recommendation parties with similar tastes will be higher. Therefore, compared with generating recommendation information for a recommendation party only based on the historical behavior of the recommendation party itself, generating recommendation information based on the similarity between recommendation direction vectors can more effectively mine the historical behavior information of similar recommendation parties and expand the range of products in the recommendation information. Similarly, the similarity and relevance of the recommended items recommended by recommendation parties with similar tastes are also higher. Therefore, compared with generating recommendation information for a purchaser only based on the historical behavior of the purchaser, generating recommendation information based on the similarity of recommended item vectors can also provide more product options for the purchaser. In addition, if the similarity between the recommendation direction vector and the recommended item vector is high, it indicates that a certain recommendation party is very likely to recommend this recommended item. Therefore, in some examples, recommendation information can also be generated for the recommendation party based on the similarity between the recommendation direction vector and the recommended item vector to assist the recommendation party in product selection.

[0063] Optionally, the above process of generating the recommendation information to be delivered adopts a data mining technique based on a graph model, and further discovers the similarity between multiple recommendation parties, the similarity or relevance between multiple recommended items, and the relevance between multiple recommendation parties and recommended items. Utilizing graph data mining technology is an important direction in the fields of computer science and artificial intelligence. Based on the data mining technology, the embodiments of the present disclosure can analyze the associations among recommendation parties, purchasers, and recommended items according to the conversion relationship graph, and flexibly generate personalized recommendation information for the purchaser or the recommendation party (influencer). The present disclosure will hereinafter refer to Figure 13 the process of retrieving similar recommendation direction vectors / recommended item vectors is described, which will not be elaborated here in the present disclosure.

[0064] Continue to refer to Figure 3 , the promotion platform installed on Figure 1 the server 110 can deliver the recommendation information generated in step S240 to different terminals 120 according to different scenarios, so as to facilitate the recommendation party or the purchaser to view the recommendation information. And the recommendation party can further select products according to the delivered recommendation information, and then generate more recommendation data.

[0065] As described above, by using the method, apparatus, device, computer-readable storage medium, and computer program product for generating recommendation information according to the various aspects of the present disclosure, it is possible to construct a conversion relationship graph between the recommender and the recommended item, and based on this conversion relationship graph, assist the recommender in product selection or the purchaser in purchasing goods, thereby expanding the product recommendation scope, improving the product selection efficiency, and promoting the increase in transaction volume and transaction efficiency. In addition, the present disclosure also improves the calculation efficiency based on large-scale graph storage technology and efficient graph calculation technology.

[0066] Next, refer to Figures 4 to 6 to further illustrate the process of obtaining the conversion data corresponding to the recommender's recommendation of the recommended item in step S210. Figure 4 FIG. is an example diagram showing part of the transaction data according to an embodiment of the present disclosure. Figure 5 FIG. is an example diagram showing part of the recommendation data according to an embodiment of the present disclosure. Figure 6 FIG. is an example diagram showing part of the conversion data according to an embodiment of the present disclosure.

[0067] As an example, step S210 further includes: aggregating the recommendation data of the recommender's recommendation of the recommended item and the transaction data of the purchaser's purchase of the recommended item to generate the conversion data corresponding to the recommender's recommendation of the recommended item, where the purchaser purchases the recommended item at least partially based on the recommendation information of the recommender for the recommended item.

[0068] For example, as described in Figure 3 , the recommendation data of the recommender's recommendation of the recommended item is generated based on the behavior of the recommender's product selection, and the transaction data of the purchaser's purchase of the recommended item is generated based on the purchase behavior. The recommendation data and the transaction data will be stored in a database or a data warehouse. Those skilled in the art should understand that the conversion data generated by aggregation can be generated based on only part of the recommendation data and transaction data, without traversing / statistical all the recommendation data and transaction data.

[0069] Optionally, refer to Figure 5 , the above aggregation process further includes: based on the recommendation data of the recommender's recommendation of the recommended item, obtaining multiple recommendation records, each recommendation record including a recommender identifier, a recommended item identifier, and a recommendation event identifier. Figure 5 Exemplarily shows two recommendation records in the recommendation data. In recommendation event E0, recommender K0 shared recommended item P0. In recommendation event E1, recommender K1 also shared recommended item P0.

[0070] Then, optionally, refer to Figure 4 , based on the transaction data of the purchaser's purchase of the recommended item, obtaining multiple transaction records, each transaction record including a purchaser identifier, a recommended item identifier, and a recommendation event identifier.Figure 4 Three transaction records in the transaction data are exemplarily shown. In recommendation event E0, purchasers C0 and C1 purchased recommended item P0 based on the recommendation of recommender K0. In recommendation event E1, purchaser C2 purchased recommended item P0 based on the recommendation of recommender K1.

[0071] Next, optionally, referring to Figure 6 , multiple recommendation records and multiple transaction records can be aggregated with the recommendation event identifier as the aggregation dimension to generate multiple conversion records, each conversion record including a recommender identifier, a recommended item identifier, and a purchaser identifier; and integrating the multiple conversion records into the conversion data corresponding to the recommender's recommendation of the recommended item.

[0072] Figure 6 Three conversion records are exemplarily shown, which respectively represent: purchaser C0 purchased recommended item P0 based on recommender K0's recommendation of recommended item P0, purchaser C1 purchased recommended item P0 based on recommender K0's recommendation of recommended item P0, and purchaser C2 purchased recommended item P0 based on recommender K1's recommendation of recommended item P0. Thus, the conversion relationship graph can then be determined based on the conversion data corresponding to the recommender's recommendation of the recommended item.

[0073] Next, referring to Figures 7 to 10 to further illustrate the process of obtaining the conversion relationship graph in step S210. Figure 7 is an example diagram showing the relationship of recommender-recommended item-purchaser corresponding to Figure 6 according to an embodiment of the present disclosure. Figure 8 is an example diagram showing the conversion relationship graph of recommender-recommended item corresponding to Figure 6 according to an embodiment of the present disclosure. Figure 9 is a partial example diagram showing the overall business diagram according to an embodiment of the present disclosure. The conversion relationship graph can exist in various forms, Figures 7 - 9 and this is only one of several forms, and those skilled in the art should understand that the present disclosure is not limited thereto.

[0074] For example, optionally, the process of obtaining the conversion relationship graph in step S210 further includes: traversing each recommendation record in the recommendation data of the recommender's recommendation of the recommended item to determine the edge connecting the recommender node and the recommended item node in the conversion relationship graph. Traversing each transaction record in the transaction data of the purchaser's purchase of the recommended item, and based on the conversion data corresponding to the recommender's recommendation of the recommended item, determining the weight corresponding to the edge connecting the recommender node and the recommended item node. Matching the recommender node, the recommended item node, and the weight corresponding to the edge connecting the recommender node and the recommended item node to generate the basic conversion relationship unit in the conversion relationship graph.

[0075] First, according toFigures 4 to 6 The data can be used to construct Figure 7 the exemplary relationship diagram of recommender - recommended item - purchaser as shown. Refer to Figure 7 , Figure 7 There are three types of nodes and three types of edges in it. The three types of nodes are the recommender node, the recommended item node, and the purchaser node. The three types of edges are the recommendation edge, the purchase edge, and the attention edge.

[0076] For example, each recommendation record in the recommendation data of the recommender recommending the recommended item can be traversed to determine Figure 7 each of the recommendation edges as shown. For example, through the recommendation record of recommender K0 recommending recommended item P0, the recommendation edge connecting recommender K0 and recommended item P0 can be determined, which shows that recommender K0 has recommended recommended item P0. The exemplary weight of the recommendation edge can be the recommendation frequency, the recommendation intensity, etc. The recommendation edge can also not be set with a weight, and the present disclosure is not limited thereto.

[0077] Again, for example, each transaction record in the transaction data of the purchaser purchasing the recommended item can be traversed to determine Figure 7 the purchase edges as shown. For example, through the transaction record of purchaser C2 purchasing recommended item P0, the shopping edge connecting recommended item node P0 and purchaser node C2 can be determined, which shows that purchaser node C2 has purchased recommended item node P0. The exemplary weight of the purchase edge can be the purchase quantity, the transaction amount, the profit, etc. The purchase edge can also not be set with a weight, and the present disclosure is not limited thereto. Again, for example, in the case where there is a historical record of purchaser C0 following recommender K0, the attention edge connecting purchaser node C0 and recommender node K0 can also be determined by traversing the attention historical record, which shows that purchaser C0 follows recommender K0 and purchases the recommended item based on the recommendation of recommender K0. The exemplary weight of the attention edge may be the attention duration, the intimacy, etc., and the present disclosure is not limited thereto.

[0078] Next, the information in Figure 7 or Figure 6 can be further integrated / aggregated to obtain Figure 8 the exemplary recommender - recommended item relationship diagram as shown. For example, based on the conversion data corresponding to the recommender recommending the recommended item, the conversion records with the same recommended item identifier and recommender identifier can be aggregated / statistically analyzed to determine Figure 8 the weight corresponding to the edge connecting the recommender node and the recommended item node in it. The weight indicates one or a combination of the following: the value of the recommended item, the purchase quantity or the total purchase value generated based on the recommendation of the recommended item by the recommender. Figure 8 Exemplarily shows a recommendation edge with the purchase quantity as the weight, and the present disclosure is not limited thereto. Again, for example, it can also be traversed Figure 7, to determine weights for the recommended edges. The present disclosure is not limited thereto.

[0079] Figure 8 Only an exemplary transformation relationship graph including only two basic transformation relationship units is shown. Among them, the recommender node K0, the recommended item node P0, and the edge connecting the recommender node K0 and the recommended item node P0 (whose weight is 2) form a basic transformation relationship unit; the recommender node K1, the recommender node P0, and the edge connecting the recommender node K1 and the recommended item node P0 (whose weight is 1) form another basic transformation relationship unit. However, due to the huge amount of data, it is often difficult for a single processor to complete the graph construction process in the entire business scenario, and big data graph construction technology needs to be used.

[0080] Figure 9 A partial sub-graph of the complete business graph processed by the big data graph construction technology is shown, which includes multiple basic transformation relationship units. Since the accuracy and integrity of the transformation relationship graph can significantly affect the accuracy of subsequent recommended item vectors and recommender vectors, it is necessary to construct a complete transformation relationship graph as accurately as possible. Optionally, the above integrity means the ability to provide all graph data without missing nodes and edges. Accuracy means eliminating data with obvious errors to avoid causing deviations to the model. See Figure 9 , both recommenders K0 and K1 have recommended the recommended item P0. Both recommenders K1 and K2 have recommended the product P11. Thus, Figure 9 the potential associations between recommenders K0, K1, K2 and the recommended items they have recommended can be reflected. Optionally, during the big data graph construction process, richer node and edge information can also be added, such as adding information related to the attributes of the recommended item itself to the recommended item node. New nodes can also be introduced, such as the buyer or the returner, etc.

[0081] For example, the process of traversing all recommended data and transaction data in step S210 involves a large amount of aggregation calculations and edge matching calculations. To further meet the requirements of integrity, the embodiments of the present disclosure can also use map-reduce or spark or MPI for distributed computing to solve the problems of insufficient single-machine computing power and memory. The above graph construction methods and tools are used to improve the integrity of the transformation relationship graph. Another example is to further improve Figure 9 the accuracy, and filters can be used to eliminate abnormal data. Abnormal data includes recommended data caused by the over-selection behavior of some recommenders (for example, recommenders often select too many recommended items during the process of testing API interfaces) and abnormal transaction data of some recommended items (such as brushing orders). By eliminating abnormal data, the accuracy of subsequent recommender vectors and recommended item vectors can be improved. The above means are all exemplary means, and the present disclosure is not limited thereto.

[0082] However, since the conversion data is massive and changes in real time, and the conversion relationship graph may be divided into multiple sub-graphs and stored in a distributed storage. The conversion relationship graph will be updated along with the update of the recommendation data and transaction data in the database. Therefore, the embodiments of the present disclosure can optionally also use a distributed graph storage engine to implement the storage and update of the above conversion relationship graph data.

[0083] For example, when a single machine cannot store all the data, a distributed architecture can be used to split the large graph to store different subgraphs in a distributed manner. According to the throughput and cost requirements of each storage device / memory / storage segment used to store the data corresponding to different subgraphs, different storage devices / memory / storage segments can choose different storage engines. For example, a key-value (KV) storage engine based on memory, SSD or HDD can be used. For these storage engines, they use the storage method of key-value pairs to store the above-mentioned basic transformation relationship units.

[0084] Different subgraph partitioning methods will affect the values corresponding to the key-value pairs in the above storage method. As an example of a subgraph partitioning method based on nodes, for the recommender, the recommender identifier can be used as the key, and the value is the set of all edges connecting the recommender. Each edge in the set has the weight of the corresponding edge in the basic conversion unit and the recommender identifier. As an example, for the recommended item, the recommender identifier can be used as the key, and the value is the set of all edges connecting the recommended item. Each edge in the set has the weight of the corresponding edge in the basic conversion unit and the recommender identifier. In addition, the subgraph partitioning can also be performed based on the edges, and the embodiments of the present disclosure do not limit the partitioning method of the subgraph of the conversion relationship graph.

[0085] Optionally, the present disclosure further optimizes the consistency and fault tolerance of storage. For example, the above-mentioned multiple storage devices / memories / storage segments can form two data shards. The two data shards store different data. Optionally, a unified gateway can be used to receive requests to update the transformation relationship graph or store a new transformation relationship graph, and then the gateway forwards the requests to different data shards according to the data in the request. In addition, in order to improve fault tolerance, each data shard may also include a spare storage device / memory / storage segment for storing spare data. In order to improve consistency, the raft or paxos algorithm can also be used to select the master node for isomorphic key-value pairs (for example, all with the recommender as the key, or all with the recommended item as the key).

[0086] The following will refer to Figure 10 An example of storing and updating a transformation relationship graph is further described. Figure 10 A flowchart of storing and updating a transformation relationship graph according to an embodiment of the present disclosure is shown.

[0087] See alsoFigure 10 , the above step S210 further includes: obtaining a first basic transformation relation unit that is not incorporated into the transformation relation graph, and determining a data shard corresponding to the first basic transformation relation unit, where the data shard is used to store a sub-graph of the transformation relation graph. Updating the corresponding data shard by using the first basic transformation relation unit. For example, Figure 10 shows an example first basic transformation relation unit (K0, P1, 7) that is not incorporated into the transformation relation graph. The gateway will query the data shard corresponding to the first basic transformation relation unit. As a simple example, data shard A includes a sub-graph related to recommenders with even recommender identifiers, and data shard B includes a sub-graph related to recommenders with odd recommender identifiers. According to the above rules, the gateway forwards the first basic transformation relation unit (K0, P1, 7) to data shard A to update the sub-graph stored in data shard A.

[0088] Optionally, the updating the corresponding data shard by using the first basic transformation relation unit further includes: obtaining the recommender identifier, the recommended item identifier, and the weight corresponding to the edge connecting the recommender node and the recommended item node in the first basic transformation relation unit; based on the identifier of the recommender and the identifier of the recommended item in the first basic transformation relation unit, determining whether the sub-graph of the transformation relation graph stored in the data shard includes a second basic transformation relation unit with the same identifiers of the recommender and the recommended item as those in the first basic transformation relation unit; in the case where the sub-graph of the transformation relation graph stored in the data shard includes the second basic transformation relation unit, updating the weight in the second basic transformation relation unit based on the weight in the first basic transformation relation unit; and in the case where the sub-graph of the transformation relation graph stored in the data shard does not include the second basic transformation relation unit, storing the first basic transformation relation unit.

[0089] Continuing with Figure 10 as an example, the main node of data shard A receives a request to write the first basic transformation relation unit (K0, P1, 7). The main node A determines whether there is a recommender node with the recommender identifier K0 in the sub-graph stored in data shard A. As Figure 10As shown, in data shard A, there is a recommender node with recommender identifier K0, and there is also an edge connecting this recommender node K0 and item node P1, and the weight corresponding to this edge is 6. Then, according to the first basic transformation relation unit (K0, P1, 7), data shard A updates the weight of the edge connecting recommender node K0 and item node P0 in this subgraph to 7 + 8 = 15. For example, assume there is another basic transformation relation unit (K0, P10, 1) to be written. Since there is no edge connecting recommender node K0 and item node P10 in the subgraph on data shard A, data shard A will add an edge connecting recommender node K0 and item node P10 in the subgraph, and the weight corresponding to the edge is 1.

[0090] For another example, assume there is yet another basic transformation relation unit (K101, P1, 12) to be written, and there is no recommender node with recommender identifier K101 on data shard A. Data shard A will create a new recommender node K101 and create an edge connecting recommender K101 and item P1, and the weight of this edge is 12. The above uses data shard A as an example for illustration. Data shard B will also store / update the subgraph data on it in a similar manner, which is not elaborated herein in the present disclosure. Those skilled in the art should understand that although only two data shards are shown, the present disclosure may correspondingly include more data shards, and may correspondingly adopt different ways to store the transformation relation graph distributively. The present disclosure does not limit this.

[0091] Next, refer to Figure 11 to further describe the process of sampling the transformation relation graph to generate a set of recommendation relation sequences. Figure 11 FIG. is an example diagram showing the determination of a set of recommendation relation sequences according to an embodiment of the present disclosure.

[0092] In order to extract the spatial information (such as the association relationship and similarity with other nodes in space) corresponding to each node in the transformation relation graph to obtain the representation vector of each node, it is usually necessary to traverse the transformation relation graph. However, the transformation relation graph includes a huge number of nodes and edges, and the transformation relation graph may also change in real time according to the behaviors of recommenders and purchasers. It is almost impossible to completely traverse all the nodes and edges in the transformation relation graph. Therefore, it is necessary to sample the data in the transformation relation graph to save the amount of calculation. To further improve the operation efficiency, a sampling method combined with batch processing technology can also be adopted. The sampling methods include Struc2Vec, DeepWalk, Node2Vec, etc. Considering the computational complexity and algorithm flexibility, Node2Vec is more suitable for the embodiments of the present disclosure. The following takes Figure 11 the sampling in the Node2Vec manner in as an example for illustration. Those skilled in the art should understand that the present disclosure is not limited thereto.

[0093] For example, the above step S220 may further include: based on the conversion relationship graph, determining a starting element corresponding to each recommendation relationship sequence in the recommendation relationship sequence set, where the starting element is a recommender identifier or a recommended item identifier; and starting from the starting element corresponding to each recommendation relationship sequence, randomly walking to generate each recommendation relationship sequence in the recommendation relationship sequence set, and the lengths of all the recommendation relationship sequences are less than a preset maximum length.

[0094] For example, Figure 11 Fig. shows a sub-graph example in the conversion relationship graph, which includes recommender nodes K0, K3, K2, and K8, and recommended item nodes P1, P2, P3, P4, P7, and P9, all of which can be used as starting elements in the recommendation relationship sequence. Therefore, a recommender starting point list with the recommender node as the starting element can be constructed. Of course, a corresponding recommended item starting point list can also be constructed, and the present disclosure is not limited thereto. Each starting element can correspond to multiple different recommendation relationship sequences, which depends on the connection situation between the starting element and other nodes in the sub-graph, and the present disclosure does not limit this. The following takes the recommender node K0 as the starting element as an example to illustrate how to perform random walking to obtain the recommendation relationship sequence corresponding to this node, and the recommendation relationship sequences corresponding to other nodes can be obtained similarly.

[0095] As Figure 11 shown, starting from the recommender node K0, there can be multiple walking paths. For example, the recommendation relationship sequences [K0, P4, K3, P9, K8], [K0, P2, K3, P9, K2], etc. can be correspondingly generated. Among them, in each of the recommendation relationship sequences, the recommender corresponding to the recommender identifier recommends the recommended item corresponding to the recommended item identifier adjacent to the recommender identifier. As described above, in an actual application scenario, traversing all paths starting from K0 may consume too much computing power, and it is necessary to randomly generate part of the recommendation relationship sequences according to certain rules. Such a process is called random walking.

[0096] For example, the random walk can be performed according to the following rules, and those skilled in the art should understand that the present disclosure is not limited to this. As an example, the random walk generates each recommendation relationship sequence in the recommendation relationship sequence set and further includes: for each recommendation relationship sequence, adding the starting element corresponding to the recommendation relationship sequence to the recommendation relationship sequence, and using the starting element as the current element, and determining the next element of the current element based on the depth walk parameter, the breadth walk parameter, and the weights corresponding to each edge connecting the current element, the next element is different from the element already added to the recommendation relationship sequence; and adding the next element to the recommendation relationship sequence and using the next element as the current element until the length of the recommendation relationship sequence is equal to the preset maximum length or the next element is empty.

[0097] Among them, the depth walk parameter and the breadth walk parameter are determined according to business needs. If the depth walk parameter is high, the sequence generated by the random walk will more reflect the structure of the conversion relationship graph. For example, if the depth walk parameter is high, starting from the recommender node K0, it will most likely walk in the direction indicated by the solid arrow to obtain the recommendation relationship sequence [K0, P4, K3, P9]. If the breadth walk parameter is high, the sequence generated by the random walk will more reflect the homogeneity of the conversion relationship graph. See Figure 11 For example, if the breadth walk parameter is high, starting from the recommender node K0, it will most likely walk along the direction indicated by the dotted arrow, obtaining the recommendation relationship sequence [K0, P1], [K0, P2, K3], etc. In addition, if the weight of the edge connecting the recommender node K0 and the recommended object node P4 is higher than other edges connecting the recommender node K0, then during the random walk, it is most likely to choose to walk from the recommender node K0 to the recommended object node P4.

[0098] For example, for each recommended relationship sequence, the walk sampling engine will maintain two data: the current node and the generated sequence. When the current node is the starting element, the length of the generated sequence is 1. For example, suppose a recommended relationship sequence is generated for K0. At the initial moment, the current node is K0, the recommended relationship sequence is [K0], and the length is 1. At this time, based on the depth walk parameter, the breadth walk parameter, and the weights corresponding to each edge connecting K0, the walk sampling engine will determine the next element to go to.

[0099] As an example, the process of determining the next element when the current node is K0 is briefly introduced below. The neighbor nodes of K0 are P1, P2, P3, and P4, all of which may be the next node to be traversed to. At this time, during the first traversal, the probability of moving towards each neighbor node can be determined only based on the weights of the respective edges connecting K0. For example, the probability of moving towards P1 is 90 / 640, the probability of moving towards P2 is 400 / 640, the probability of moving towards P3 is 50 / 640, and the probability of moving towards P4 is 100 / 640. The traversal sampling engine will randomly select the next element of K0 with such a probability distribution.

[0100] Suppose the next element is determined to be P4. At this time, P4 is added to the tail of the recommended relationship sequence, and the recommended relationship sequence is updated to [K0, P4], with a length of 2. Then, P4 will be used as the current element to continue the next traversal. Suppose the breadth traversal parameter is p and the depth traversal parameter is q. p controls the probability of being able to return to the neighbors (P1, P3, P2) of K0 after two traversals. q controls the probability of being able to move towards the non-neighbors of K0 after two traversals. In other words, p is the probability of controlling the return; q is the probability of controlling the outward movement. As Figure 11 shown, if moving from P4 to K3, then it is still possible to return to the neighbors P2 and P3 of K0. If moving from P4 to K2, then it is less likely to return to any neighbor node of K0. If the p value is much higher than the q value, then most likely the node K3 will be selected, otherwise, most likely the node K2 will be selected. The above probabilities can also be combined with the weights of the edges to determine the probabilities of moving towards each neighbor node. For example, assume p = 2 and q = 0.1. Then the probability of moving from P4 to K3 is 80 * 0.1 / (80 * 0.1 + 77 * 2) = 0.049; then the probability of moving from P4 to K2 is 77 * 2 / (80 * 0.1 + 77 * 2) = 0.951. The above scheme is only an example, and those skilled in the art should understand that the present disclosure is not limited thereto.

[0101] Then the traversal engine will continue to traverse until the length of the recommended relationship sequence reaches a preset maximum length, for example, 30. Or, until the next element is empty. In this way, a recommended relationship sequence including multiple alternating recommender identifiers and recommended item identifiers can be generated. The traversal engine will summarize all the recommended relationship sequences generated with the starting element being K0 and store them in the recommended relationship sequence set. The above traversal process is only an example, and those skilled in the art should understand that other methods can also be used to obtain the recommended relationship sequence set, and the present disclosure is not limited thereto.

[0102] Next, refer to Figure 12 to further describe the process of obtaining the recommendation direction vector and the recommended item vector using the recommended relationship sequence set. Figure 12 is an example diagram showing the process of training a word vector conversion model according to an embodiment of the present disclosure.

[0103] As an example, step S230 further includes: for each recommender node or each recommended item node in the conversion relationship graph, using a word vector conversion model to generate a recommender vector corresponding to the recommender or a recommended item vector corresponding to the recommended item, wherein the word vector conversion model is trained by the set of recommendation relationship sequences, and the distance between two recommender vectors with higher similarity is smaller, and the distance between two recommended item vectors with higher similarity is smaller.

[0104] As Figure 12 shown, after obtaining the set of recommendation relationship sequences, the learning of the recommender vector or the recommended item vector is performed through the sampling-classification process of the Skip-Gram neural network model or the CBOW neural network model. Hereinafter, the Skip-Gram neural network model is taken as an example of the word vector conversion model for illustration, and the present disclosure is not limited thereto.

[0105] As an example, the goal of the Skip-Gram neural network model is to predict the probability of occurrence of other recommender identifiers or recommended item identifiers in the context of the recommendation relationship sequence given a certain recommender identifier or recommended item identifier. Among them, the Skip-Gram neural network model generally includes an input layer, a hidden layer, and an output layer. The input layer inputs a certain recommender identifier or recommended item identifier, and the output layer outputs the probability of the adjacent recommender identifier or recommended item identifier of the recommender identifier or recommended item identifier. The Skip-Gram neural network model will train the center word vector matrix of the training dictionary and the context vector matrix of the dictionary, but the goal of the Skip-Gram neural network model is actually only to learn the weight matrix of the hidden layer, and this weight matrix can actually also be understood as the initial vector of the network representation learning to be learned. Therefore, after training, the output layer can be discarded during the inference process.

[0106] As an example, the training of the word vector conversion model includes: using each recommender identifier and each recommended item identifier in the conversion relationship graph as the center word respectively, calculating the co-occurrence vector matrix corresponding to the center word based on the set of recommendation relationship sequences, the co-occurrence vector indicating the conditional probability distribution of the corresponding relationship between the center word and the neighborhood of the center word in the conversion relationship graph, determining the predicted co-occurrence vector matrix corresponding to the center word based on the center word, using the center word vector matrix of the dictionary and the context word vector matrix in the word vector conversion model, and adjusting the center word vector matrix of the dictionary and the context word vector matrix of the dictionary so that the predicted co-occurrence vector matrix corresponding to the center word approaches the co-occurrence vector matrix corresponding to the center word.

[0107] For example, the above-mentioned shared vector matrix is the probability that other recommender identifiers / item identifiers appear within the context of 2M words from the center word, which is the recommender identifier / item identifier serving as the center word, in the set of recommendation relationship sequences, where M is an integer greater than zero. In each iteration, the Skip-gram neural network model will specify 2M context words of a center word to train the dictionary center word vector matrix and the dictionary context word vector matrix corresponding to this center word. In the next iteration, the next center word in the recommendation relationship sequence will be specified, and its 2M context words will be examined for training.

[0108] For example, the recommender identifier and / or item identifier can first be converted into its one-hot representation, which is a V*1-dimensional vector. The one-hot representation is also known as a one-hot vector (also called a one-hot vector). Only one bit is allowed to be 1 in the one-hot vector. Then, this one-hot representation can be multiplied by a d*V-dimensional dictionary center word vector matrix to be trained. The dictionary center word vector matrix is composed of the respective recommender vectors and / or item vectors in the dictionary. Each column of the dictionary center word vector matrix is a recommender vector and / or item vector in the dictionary. The dictionary contains d-dimensional abstract information for each recommender / item, indicating the corresponding relationship information / similarity information between recommenders / items that may appear in the context when a certain recommender identifier / item identifier serves as the center word.

[0109] After the multiplication of the one-hot identifier and the dictionary center word vector matrix, a d-dimensional vector will be obtained. This d-dimensional vector will be multiplied by the V*d-dimensional dictionary context vector matrix to be trained, resulting in multiple V-dimensional vectors. The abstract information stored in the d-dimensional vectors in the dictionary context vector matrix refers to the relationship between the recommender identifier / item identifier as the word in the context and the recommender identifier / item identifier as the center word. These multiple V-dimensional vectors will pass through an output layer (which includes, for example, a softmax layer and a normalization layer) to obtain the predicted co-occurrence vector matrix corresponding to the center word.

[0110] The training process is to continuously adjust the parameters in the above-mentioned dictionary center word vector matrix and the dictionary context word vector matrix until the predicted co-occurrence vector matrix corresponding to the center word approaches the co-occurrence vector matrix corresponding to the center word.

[0111] After training, the word vector conversion model will take the recommender identifier / item identifier as the input and output the recommender vector / item vector. Among them, the smaller the distance between two recommender vectors with higher similarity, and the smaller the distance between two item vectors with higher similarity.

[0112] Although referring toFigure 11 and Figure 12 A solution for calculating a recommended direction vector / recommended item vector using a transformation relationship graph is described in detail. However, the method based on the graph model is also applicable to this disclosure, including but not limited to GNN (Graph Neural Network, such as GraphSage, PinSage), and neural network optimization solutions such as introducing an attention mechanism can also be used to further improve the accuracy of the graph model. This disclosure is not limited thereto.

[0113] Next, refer to Figure 13 to further describe the process of retrieving recommended direction vectors and recommended item vectors with high similarity using a similarity retrieval scheme. Figure 13 is an example diagram showing the process of retrieving similar recommended direction vectors and recommended item vectors according to an embodiment of this disclosure.

[0114] The recommended direction vectors and recommended item vectors obtained with reference to the above description may be massive. And the smaller the distance between two vectors, the higher the similarity between the two vectors. However, to find the top N vectors with the highest similarity to a certain recommended direction vector / recommended item vector, comparing the recommended direction vector / recommended item vector with all vectors in the database is unacceptable in terms of computing time and computational complexity. For this reason, in order to further improve the operation efficiency, this disclosure can also adopt the following vector approximate indexing scheme to quickly provide similar vectors. The vector approximate indexing scheme includes: Faiss, HNSW, SPTAG, SCANN, etc. This disclosure does not limit the specific vector approximate indexing scheme.

[0115] As an example, referring to Figure 13 , step S240 further includes: determining a recommended party identifier or a recommended item identifier based on the scenario information corresponding to the recommended information to be delivered; determining a recommended direction vector or a recommended item vector corresponding to the recommended party identifier or the recommended item identifier, and using the recommended direction vector or the recommended item vector as a query key; using the query key to retrieve multiple vectors similar to the recommended direction vector or the recommended item vector from the retrieval database, where the multiple vectors are a combination of multiple recommended direction vectors, multiple recommended item vectors, or recommended direction vectors and recommended item vectors, and generating the recommended information to be delivered based on the multiple vectors.

[0116] For example, the scenario information may indicate delivering recommended information for assisting product selection to the recommended party, or may indicate delivering recommended items to the purchasing party. The following introduces 5 possible example scenarios, and those skilled in the art should understand that this disclosure is not limited thereto.

[0117] Example 1: When the scenario information indicates recommending alternative recommended items to the buyer based on a key recommender, the recommender identifier in step S240 is the recommender identifier corresponding to the key recommender. Then, based on the recommender vector corresponding to the key recommender identifier, the first recommended information to be delivered can be generated, and the first recommended information includes information about multiple alternative recommended items associated with the key recommender. Since the recommender vector and the recommended item vector are obtained based on the same transformation relationship graph, it is possible to directly obtain the recommended items that the key recommender may like according to the recommender vector corresponding to the key recommender identifier. After obtaining such recommended item information, the recommended information can be directly delivered to the buyer.

[0118] Example 2: When the scenario information indicates recommending alternative recommended items to a key recommender, the recommender identifier in step S240 is the recommender identifier corresponding to the key recommender. Then, based on the recommender vector corresponding to the key recommender identifier, the second recommended information to be delivered is generated, and the second recommended information includes information about multiple alternative recommenders similar to the key recommender. In such a scenario, it is possible to recommend to influencer A influencers B who are similar to A in terms of product selection style and transaction results. Thus, influencer A may follow influencer B and assist in product selection based on the recommended items corresponding to influencer B.

[0119] Example 3: When the scenario information indicates recommending alternative recommended items to a key recommender, the recommender identifier in step S240 is the recommender identifier corresponding to the key recommender. Then, based on the recommender vector corresponding to the key recommender, the third recommended information to be delivered is generated, and the third recommended information includes information about multiple alternative recommended items associated with the key recommender. In such a scenario, it is possible to directly recommend to influencer A the recommended items that have been recommended by influencer B, who is similar to A in terms of product selection style and transaction results. Thus, influencer A can directly assist in product selection based on the recommended items corresponding to influencer B.

[0120] Example 4: When the scenario information indicates recommending alternative recommended items to the buyer based on the recommended items the buyer has previously purchased, the recommended item identifier in step S240 is the recommended item identifier corresponding to the recommended items the buyer has previously purchased. Then, based on the recommended item vector corresponding to the recommended items the buyer has previously purchased, the fourth recommended information to be delivered can be generated, and the fourth recommended information includes information about multiple alternative recommended items associated with the recommended items the buyer has previously purchased. This scenario can be used for real-time recommendation, which recommends approximate items of the products that the buyer has recently been interested in to the buyer.

[0121] Example 5: In the case where the scenario information indicates recommending alternative recommended items to the purchaser or the key recommender based on the currently browsed recommended item, the recommended item identifier in step S240 is the recommended item identifier corresponding to the currently browsed recommended item. Then, based on the recommended item vector corresponding to the currently browsed recommended item, the fifth recommended information to be delivered is generated, and the fifth recommended information includes information on multiple alternative recommended items associated with the currently browsed recommended item. This scenario usually corresponds to the "You may like" function at the bottom of the product details page to recommend similar products to the currently browsed product.

[0122] For example, then refer to Figure 13 , the retrieval database includes at least one of a plurality of recommender entries and a plurality of recommended item entries. Wherein, each recommender entry includes: a recommender identifier and a recommender vector corresponding to the recommender identifier, and each recommended item entry in the retrieval database includes: a recommended item identifier and a recommended item vector corresponding to the recommended item identifier. The retrieval database includes at least one of a recommender retrieval database and a recommended item retrieval database. Note that although the recommender retrieval database and the recommended item retrieval database are separately shown in Figure 13 , those skilled in the art should understand that the recommender retrieval database and the recommended item retrieval database can actually be different components of the same retrieval database as long as they can store the corresponding information.

[0123] Optionally, the retrieval database can be an Approximate Nearest Neighbor (ANN) retrieval database. In the approximate nearest neighbor retrieval library, the approximate nearest neighbor retrieval method can be adopted to quickly find one or more data similar to the query key. Specifically, the approximate nearest neighbor retrieval can utilize the characteristics of the cluster-like aggregation distribution formed between massive data, and classify or encode the recommender vectors / recommended item vectors in the retrieval database by analyzing and clustering the data. Then, the approximate nearest neighbor retrieval can predict the data category to which the query key belongs and return some or all of the category as the retrieval result. Optionally, the retrieval database can be constructed as an approximate nearest neighbor retrieval library using tools such as Annoy and Faiss. Of course, other tools can also be used to construct the approximate nearest neighbor retrieval library, and the present disclosure does not limit this.

[0124] Thus, based on the above retrieval scheme, the similarity between multiple recommenders, the similarity or relevance between multiple recommended items, and the relevance between multiple recommenders and recommended items can be mined. As described above, different recommended information can be further generated according to the different scenario information of the application scenario.

[0125] By using the method, apparatus, device, computer-readable storage medium, and computer program product for generating recommendation information according to the various aspects of the present disclosure above, it is possible to construct a conversion relationship graph between the recommender and the recommended item, and based on this conversion relationship graph, assist the recommender in product selection or the purchaser in purchasing goods, thereby expanding the product recommendation scope, improving the product selection efficiency, and promoting the increase in the transaction volume and transaction efficiency. In addition, the present disclosure also improves the calculation efficiency based on large-scale graph storage technology and efficient graph calculation technology.

[0126] In addition, the device according to the embodiments of the present disclosure (for example, the recommendation information processing device, the recommendation information sorting device, etc.) can also be implemented by means of Figure 14 the architecture of the exemplary computing device shown. Figure 14 FIG. shows a schematic diagram of the architecture of an exemplary computing device according to an embodiment of the present disclosure.

[0127] As Figure 14 shown, the computing device 1100 may include a bus 1110, one or more CPUs 1120, a read-only memory (ROM) 1130, a random access memory (RAM) 1140, a communication port 1150 connected to a network, an input / output component 1160, a hard disk 1170, etc. The storage device in the computing device 1100, such as the ROM 1130 or the hard disk 1170, may store various data or files used for computer processing and / or communication, as well as program instructions executed by the CPU. The computing device 1100 may also include a user interface 1180. Of course, Figure 11 the architecture shown is only exemplary. When implementing different devices, one or more components in the Figure 11 shown computing device may be omitted according to actual needs. The device according to the embodiments of the present disclosure may be configured to execute the recommendation information processing method and the recommendation information sorting method according to the various embodiments of the present disclosure above, or be used to implement the recommendation information processing device and the recommendation information sorting device according to the various embodiments of the present disclosure above.

[0128] The embodiments of the present disclosure may also be implemented as a computer-readable storage medium. Computer-readable instructions are stored on the computer-readable storage medium according to the embodiments of the present disclosure. When the computer-readable instructions are run by a processor, the recommendation information processing method and the recommendation information sorting method according to the embodiments of the present disclosure described with reference to the above drawings may be executed. The computer-readable storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory, etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0129] According to an embodiment of the present disclosure, there is also provided a computer program product or a computer program. The computer program product or the computer program includes computer-readable instructions, and the computer-readable instructions are stored in a computer-readable storage medium. A processor of a computer device may read the computer-readable instructions from the computer-readable storage medium, and the processor executes the computer-readable instructions, so that the computer device executes the methods described in the foregoing various embodiments.

[0130] In addition, according to an embodiment of the present disclosure, there is also provided a device for generating recommendation information, including: an offline module configured to: determine a conversion relationship graph based on conversion data corresponding to a recommended item recommended by a recommender, the conversion relationship graph including a plurality of recommenders and a plurality of recommended items, and an edge connecting a recommender node and a recommended item node indicating that the recommender recommends the recommended item; determine a set of recommendation relationship sequences including a plurality of recommendation relationship sequences based on the conversion relationship graph, each recommendation relationship sequence including a plurality of alternating recommender identifiers and recommended item identifiers; generate at least one of a recommender vector and a recommended item vector based on the set of recommendation relationship sequences; and an online module configured to generate recommendation information to be delivered based on the similarity between at least one of the recommender vector and the recommended item vector.

[0131] Those skilled in the art can understand that the content disclosed in the present disclosure can have various variations and improvements. For example, the various devices or components described above can be implemented by hardware, or can be implemented by software, firmware, or some or all of the combinations of the three.

[0132] In addition, as shown in the present disclosure and the claims, unless the context clearly indicates an exception, words such as "a", "an", "one", and / or "the" are not specifically singular and may also include plural. The terms "first", "second", and similar terms used in the present disclosure do not indicate any order, quantity, or importance, but are only used to distinguish different components. Similarly, words such as "including" or "comprising" mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. The terms "connected" or "coupled" and the like are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0133] In addition, flowcharts are used in the present disclosure to illustrate the operations performed by the systems according to the embodiments of the present disclosure. It should be understood that the operations before or below do not necessarily need to be executed precisely in order. On the contrary, they can be executed in reverse order or simultaneously. At the same time, other operations can also be superimposed on these processes, or one or several steps of operations can be removed from these processes.

[0134] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. It should also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and should not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0135] The foregoing has described the present disclosure in detail. However, for those skilled in the art, it is obvious that the present disclosure is not limited to the embodiments described in this specification. The present disclosure can be implemented in the form of modifications and changes without departing from the spirit and scope of the present disclosure determined by the claims. Therefore, the description in this specification is for the purpose of illustration and has no limiting significance for the present disclosure.

Claims

1. A method for generating recommendation information, comprising: Determining a conversion relationship graph based on conversion data corresponding to a recommendation object recommended by a recommender, wherein the conversion relationship graph includes a plurality of recommender nodes and a plurality of recommendation object nodes, and an edge connecting a recommender node and a recommendation object node indicates that the recommender recommends the recommendation object. The recommender node is identified by a recommender identifier, the recommendation object node is identified by a recommendation object identifier, and a weight corresponding to an edge connecting the recommender node and the recommendation object node indicates the number of purchases or the total purchase value generated based on the recommendation of the recommendation object by the recommender; Determining a set of recommendation relationship sequences including a plurality of recommendation relationship sequences based on the conversion relationship graph, each recommendation relationship sequence including a plurality of alternating recommender identifiers and recommendation object identifiers. Wherein, in each recommendation relationship sequence, the recommender corresponding to the recommender identifier recommends the recommendation object corresponding to the recommendation object identifier adjacent to the recommender identifier; Generating at least one of a recommender vector and a recommendation object vector based on the set of recommendation relationship sequences; and Generating recommendation information to be delivered based on the similarity between at least one of the recommender vector and the recommendation object vector, Wherein, the generating of the recommendation information to be delivered based on the similarity between at least one of the recommender vector and the recommendation object vector further includes: determining a recommender identifier or a recommendation object identifier based on the scenario information corresponding to the recommendation information to be delivered; Wherein, in the case where the scenario information indicates recommending an alternative recommendation object to a key recommender, the recommender identifier is the recommender identifier corresponding to the key recommender, and based on the recommender vector corresponding to the key recommender, generating the recommendation information to be delivered, and the recommendation information includes information on a plurality of alternative recommendation objects recommended by alternative recommenders similar to the key recommender.

2. The method according to claim 1, wherein, The conversion relationship graph includes a plurality of basic conversion relationship units, and each basic conversion relationship unit at least includes: an identifier of a recommender, an identifier of a recommendation object, and a weight corresponding to an edge connecting the recommender node and the recommendation object node.

3. The method according to claim 2, wherein, The determining of the conversion relationship graph based on the conversion data corresponding to the recommendation object recommended by the recommender further includes: Aggregating the recommendation data of the recommender recommending the recommendation object and the transaction data of the purchaser purchasing the recommendation object to generate the conversion data corresponding to the recommender recommending the recommendation object, wherein the purchaser purchases the recommendation object at least partially based on the recommendation information of the recommendation object by the recommender; Determining the conversion relationship graph based on the conversion data corresponding to the recommender recommending the recommendation object.

4. The method according to claim 3, wherein, The aggregating of the recommendation data of the recommender recommending the recommendation object and the transaction data of the purchaser purchasing the recommendation object further includes: Based on the recommendation data of the recommender recommending the recommendation object, obtaining a plurality of recommendation records, each recommendation record including a recommender identifier, a recommendation object identifier, and a recommendation event identifier; Based on the transaction data of the purchaser purchasing the recommendation object, obtaining a plurality of transaction records, each transaction record including a purchaser identifier, a recommendation object identifier, and a recommendation event identifier; Aggregate multiple recommendation records and multiple transaction records with the recommended event identifier as the aggregation dimension to generate multiple conversion records, and each conversion record includes a recommender identifier, a recommended item identifier, and a purchaser identifier; and Integrate the multiple conversion records into the conversion data corresponding to the recommender's recommendation of the recommended item.

5. The method according to claim 3, wherein The determining of the conversion relationship graph based on the conversion data corresponding to the recommender's recommendation of the recommended item further includes:[[]] Traverse each recommendation record in the recommendation data of the recommender's recommendation of the recommended item, and determine the edge connecting the recommender node and the recommended item node in the conversion relationship graph; Traverse each transaction record in the transaction data of the purchaser's purchase of the recommended item, and based on the conversion data corresponding to the recommender's recommendation of the recommended item, determine the weight corresponding to the edge connecting the recommender node and the recommended item node; Match the recommender node, the recommended item node, and the weight corresponding to the edge connecting the recommender node and the recommended item node to generate the basic conversion relationship unit in the conversion relationship graph.

6. The method according to claim 2, wherein, The determining of the conversion relationship graph based on the conversion data corresponding to the recommender's recommendation of the recommended item further includes:[[]] Obtain the first basic conversion relationship unit not incorporated into the conversion relationship graph, and determine the data shard corresponding to the first basic conversion relationship unit, where the data shard is used to store the sub-graph of the conversion relationship graph; Update the corresponding data shard using the first basic conversion relationship unit.

7. The method according to claim 6, wherein The updating of the corresponding data shard using the first basic conversion relationship unit further includes:[[]] Obtain the recommender identifier, the recommended item identifier, and the weight corresponding to the edge connecting the recommender node and the recommended item node in the first basic conversion relationship unit; Based on the recommender identifier and the recommended item identifier in the first basic conversion relationship unit, determine whether the sub-graph of the conversion relationship graph stored in the data shard includes a second basic conversion relationship unit with the same recommender identifier and recommended item identifier as those in the first basic conversion relationship unit; In the case where the sub-graph of the conversion relationship graph stored in the data shard includes the second basic conversion relationship unit, update the weight in the second basic conversion relationship unit based on the weight in the first basic conversion relationship unit; and In the case where the sub-graph of the conversion relationship graph stored in the data shard does not include the second basic conversion relationship unit, store the first basic conversion relationship unit.

8. The method according to claim 2, wherein, The determining of the recommendation relationship sequence set including multiple recommendation relationship sequences based on the conversion relationship graph further includes:[[]] Based on the conversion relationship graph, determine the starting element corresponding to each recommendation relationship sequence in the recommendation relationship sequence set, and the starting element is the recommender identifier or the recommended item identifier; and Starting from the starting element corresponding to each recommendation relationship sequence, randomly walk to generate each recommendation relationship sequence in the recommendation relationship sequence set, and the length of each recommendation relationship sequence is less than the preset maximum length.

9. The method according to claim 8, wherein, The randomly walking to generate each recommendation relationship sequence in the recommendation relationship sequence set further includes:[[]] For each recommendation relationship sequence, Add the starting element corresponding to the recommended relationship sequence to the recommended relationship sequence, and use the starting element as the current element. Based on the deep walk parameter, the breadth walk parameter, and the weights corresponding to the respective edges connecting the current element, determine the next element of the current element, where the next element is different from the elements that have been added to the recommended relationship sequence; and Add the next element to the recommended relationship sequence, and use the next element as the current element until the length of the recommended relationship sequence is equal to the preset maximum length or the next element is empty.

10. The method according to claim 2, wherein, The generating at least one of a recommended direction vector and a recommended item vector based on the recommended relationship sequence set further includes: For each recommended party node or each recommended item node in the conversion relationship graph, use a word vector conversion model to generate a recommended direction vector corresponding to the recommended party or a recommended item vector corresponding to the recommended item, where the word vector conversion model is trained by the recommended relationship sequence set, and the distance between two recommended direction vectors with higher similarity is smaller, and the distance between two recommended item vectors with higher similarity is smaller.

11. The method according to claim 10, wherein, The training of the word vector conversion model includes: Use the respective recommended party identifiers and the respective recommended item identifiers in the conversion relationship graph as central words, and based on the recommended relationship sequence set, calculate the co-occurrence vector matrix corresponding to the central words, where the co-occurrence vector indicates the conditional probability distribution of the corresponding relationship between the central word and the neighborhood of the central word in the conversion relationship graph. Based on the central words, use the dictionary central word vector matrix and the dictionary context word vector matrix in the word vector conversion model to determine the predicted co-occurrence vector matrix corresponding to the central words, and Adjust the dictionary central word vector matrix and the dictionary context word vector matrix so that the predicted co-occurrence vector matrix corresponding to the central words approximates the co-occurrence vector matrix corresponding to the central words.

12. The method according to claim 1, wherein The generating the recommended information to be delivered based on the similarity between at least one of the recommended direction vector and the recommended item vector further includes: Based on the recommended party identifier or the recommended item identifier, determine the recommended direction vector or the recommended item vector corresponding to the recommended party identifier or the recommended item identifier, and use the recommended direction vector or the recommended item vector as a query key; Use the query key to retrieve multiple vectors similar to the recommended direction vector or the recommended item vector from the retrieval database, where the multiple vectors are a combination of multiple recommended direction vectors, multiple recommended item vectors, or recommended direction vectors and recommended item vectors, and generate the recommended information to be delivered based on the multiple vectors; Wherein, the retrieval database includes at least one of multiple recommended party entries and multiple recommended item entries, where each recommended party entry includes: a recommended party identifier and the recommended direction vector corresponding to the recommended party identifier, and each recommended item entry in the retrieval database includes: a recommended item identifier and the recommended item vector corresponding to the recommended item identifier.

13. The method according to claim 12, wherein When the scenario information indicates recommending alternative recommended items to the purchaser based on a key recommender, the recommender identifier is the recommender identifier corresponding to the key recommender; When the scenario information indicates recommending alternative recommenders similar to the key recommender to the key recommender, the recommender identifier is the recommender identifier corresponding to the key recommender; When the scenario information indicates recommending alternative recommended items to the purchaser based on the recommended items previously purchased by the purchaser, the recommended item identifier is the recommended item identifier corresponding to the recommended items previously purchased by the purchaser; Or When the scenario information indicates recommending alternative recommended items to the purchaser or the key recommender based on the currently viewed recommended item, the recommended item identifier is the recommended item identifier corresponding to the currently viewed recommended item.

14. The method according to claim 13, wherein, Generating the recommended information to be delivered based on the multiple recommender vectors or recommended item vectors further includes one or more of the following: When the scenario information indicates recommending alternative recommended items to the purchaser based on a key recommender, generating a first recommended information to be delivered based on the recommender vector corresponding to the key recommender identifier, where the first recommended information includes information on multiple alternative recommended items associated with the key recommender; When the scenario information indicates recommending alternative recommenders similar to the key recommender to the key recommender, generating a second recommended information to be delivered based on the recommender vector corresponding to the key recommender identifier, where the second recommended information includes information on multiple alternative recommenders similar to the key recommender; When the scenario information indicates recommending alternative recommended items to the purchaser based on the recommended items previously purchased by the purchaser, generating a fourth recommended information to be delivered based on the recommended item vector corresponding to the recommended items previously purchased by the purchaser, where the fourth recommended information includes information on multiple alternative recommended items associated with the recommended items previously purchased by the purchaser; Or When the scenario information indicates recommending alternative recommended items to the purchaser or the key recommender based on the currently viewed recommended item, generating a fifth recommended information to be delivered based on the recommended item vector corresponding to the currently viewed recommended item, where the fifth recommended information includes information on multiple alternative recommended items associated with the currently viewed recommended item.

15. An apparatus for generating recommended information, comprising: One or more processors; And One or more memories, where computer-readable code is stored in the memories, and when the computer-readable code is run by the one or more processors, the one or more processors are caused to execute the method according to any one of claims 1 - 14.

Images